In trials to test the efficacy of HIV vaccines, one of the measures of interest is the effect of vaccination on viral load, or the amount of virus present in an infected person’s body, for those participants who contract HIV. But vaccine trials are long, and participants may drop out, or, if HIV-infected, start anti-retroviral therapy which affects their viral load numbers, meaning that viral load data will be missing for some subjects. Now, VIDI member Peter Gilbert and recent University of Washington PhD graduate Yuying Jin have devised a valid statistical method to deal with these missing numbers, as well as compare HIV infected people in vaccine and placebo groups in a meaningful way.
Most statistical methods to deal with this type of data analysis assume that missing data are missing completely at random (MCAR), which is an incorrect assumption in the case of HIV viral load, Gilbert said, as it is known that patients with higher viral loads are more likely to start treatment. Gilbert and Jin’s statistical method assumes that missing data depends on the characteristics of the participant, but does not depend on any unmeasured factors, a statistical assumption called “missing at random” (MAR).
Their method, which will be used to analyze the only currently ongoing vaccine efficacy trial, HIV Vaccine Trials Network trial 505, also takes into account the difficulties in comparing HIV infected people in the vaccine and placebo groups. These two groups of people are not randomized; there might be an important distinction between them, such as genetic differences or other confounding factors. While this method was devised for use in an HIV vaccine efficacy trial, it could be used for other trials examining treatment effect of a given outcome measured after participants are randomly sorted into groups.
Semiparametric estimation of the average causal effect of treatment on an outcome measured after a postrandomization event, with missing outcome data. Gilbert PB, Jin Y. Biostatistics. 2009 Oct 8.